Exponentially Modified Gaussian Distribution

The distribution of an independent Gaussian value plus a positive exponential value, yielding a precise right-skewed convolution family.

Version

v1 · 2026-10-03 · History

Domain-specific #

13214

Domain group

Formal Sciences

Origin domain

Mathematics

Subdomains

Probability Theory, Statistical Distributions → Mathematics

Aliases

Ex-Gaussian distribution, EMG distribution

Core Idea

The exponentially modified Gaussian distribution, also called the ex-Gaussian or EMG, is the law of \(Z=X+Y\) where \(X\sim N(\mu,\sigma^2)\), \(Y\sim\operatorname{Exp}(\lambda)\), \(X\) and \(Y\) are independent, and \(\sigma,\lambda>0\). The exponential variable is nonnegative; adding it to a Gaussian produces a right-skewed density. The name means a specific convolution distribution, not simply a histogram with a long right tail.^[1][2]

Writing \(\tau=1/\lambda\) for the exponential mean, the sum has \(E[Z]=\mu+\tau\) and \(\operatorname{Var}(Z)=\sigma^2+\tau^2\). Its standardized skewness is \(2\tau^3/(\sigma^2+\tau^2)^{3/2}\). The moments follow by independence and addition of cumulants; they describe the mathematical model, not proof that two physically separate processes generated any fitted dataset.^[1][3][4]

Structural Signature

Sig role-phrases: Gaussian location-and-spread component — independent positive exponential component — additive convolution — real-valued support with right skew — three-parameter moment readout — conditional empirical interpretation.

• Gaussian component: \(X\) contributes \(\mu\) and \(\sigma^2\), giving the symmetric part of the constructed law. An arbitrary symmetric component would define another family.^[1]
• Exponential component: \(Y\) has rate \(\lambda\), equivalently mean/scale \(\tau=1/\lambda\). It adds only nonnegative delay in the construction; replacing it with a gamma or lognormal term changes the family.^[1]
• Independence and sum: the density is \(f_Z=f_X*f_Y\). Independence is what licenses the ordinary convolution and the simple variance addition.^[1][3]
• Support and tail: for \(\sigma>0\) the Gaussian component has full-real support, so \(Z\) too has support over all \(\mathbb R\), even though \(Y\geq0\). The exponential addend creates the positive skew; the model does not impose a hard zero lower bound on response times.^[1]
• Readout rather than process label: \(\mu,\sigma,\tau\) summarize fitted location, symmetric spread and right-tail scale. They do not uniquely identify attention, decision or chromatographic mechanisms from shape alone.^[4]

What It Is Not

It is not any right-skewed empirical histogram. Lognormal, inverse Gaussian and mixture models can also be right-skewed, but do not have this exact independent normal-plus-exponential construction. It is not a Gaussian mixture with a random component label: EMG adds independent variates, rather than switching between component populations. The live Normal-Exponential-Gamma Distribution is a separate family, not a synonym obtained by ignoring its gamma component.^[1]

It is not the literal law \(X-Y\) for a positive exponential \(Y\): that construction has a long left tail. The frozen Wikipedia surface “Gaussian minus exponential distribution” redirects to this topic, but the sign difference is mathematically consequential; that source title is held for separate identity review and is not accepted here as an alias. Nor does a fitted ex-Gaussian \(\tau\) certify attentional lapses: original model-comparison research cautions that these parameters do not map uniquely to cognitive process variables.^[4]

Scope of Application

In chromatography, EMG functions describe some asymmetric elution peaks. Grushka's original 1972 paper explicitly treats exponentially modified Gaussian chromatographic peaks; a later original study fits EMG profiles for a homologous fatty-acid series separated by reversed-phase capillary liquid chromatography and compares fitted parameters with peak moments. Those reports support an analytical peak-shape model, not a universal claim that each tail is caused by one identifiable adsorption delay.^[5][6]

The later study also found that, for its experimental zone profiles, the empirically measured second moment did not equal the sum of fitted Gaussian and exponential variance contributions. That does not overturn the variance identity for an ideal independent sum; it limits treating fitted components as independent physical broadening sources in real chromatograms.^[6]

In response-time research, Heathcote, Popiel and Mewhort use the three-parameter ex-Gaussian to analyze Stroop reaction-time distributions, finding information about facilitation that a comparison of means alone obscured. Here the fit is often descriptive: response times are physically nonnegative, while the mathematical EMG with \(\sigma>0\) assigns some probability below zero. If that left tail or other process constraints matter, the model needs checking against alternatives rather than being treated as a literal timing mechanism.^[2][4]

Clarity

Parameter conventions differ. This entry uses exponential rate \(\lambda>0\) and scale/mean \(\tau=1/\lambda\). In SciPy's exponnorm, the shape parameter is \(K=\tau/\sigma=1/(\sigma\lambda)\), while loc is \(\mu\) and Scale is \(\sigma\). A large numerical value of \(\lambda\) therefore means a shorter exponential tail, whereas a large \(\tau\) means a longer one. Confusing rate and scale reverses interpretations.^[1]

For \(z\in\mathbb R\), one explicit density is

\[f_Z(z)=\int_0^\infty \frac{1}{\sigma\sqrt{2\pi}}e^{-(z-y-\mu)^2/(2\sigma^2)}\,\lambda e^{-\lambda y}\,dy =\frac{\lambda}{2}e^{\lambda(\mu-z)+\lambda^2\sigma^2/2}\operatorname{erfc}\!\left(\frac{\mu+\lambda\sigma^2-z}{\sqrt2\sigma}\right).\]

The integral states the construction; the error-function form is its closed-form evaluation and matches SciPy after its \(K\), loc, and Scale conversion.^[1]

Manages Complexity

The model compresses a full asymmetric density to three parameters whose contributions to the first two moments are transparent. If an observed peak broadens, one can ask whether a fitted change appears mainly in \(\sigma\) or \(\tau\); if two response-time conditions have the same mean, their distributional shapes can still differ. This is more informative than forcing every observation into a symmetric normal summary.^[6][2]

Compression has costs. Different data-generating processes can produce similar curves, and censoring or trimming valid long observations changes the tail that \(\tau\) describes. Ulrich and Miller's original truncation study warns that exclusion of extreme but valid reaction times can introduce substantial bias. A fitted component decomposition is therefore a diagnostic model, not automatic causal identification.^[3][4]

Abstract Reasoning

The convolution identity follows directly from the independent sum: for a possible observed \(z\), integrate over every nonnegative exponential contribution \(y\) and weigh the Gaussian density at \(z-y\). Independence also gives \(E[Z]=E[X]+E[Y]\) and \(\operatorname{Var}(Z)=\operatorname{Var}(X)+\operatorname{Var}(Y)\). Since the exponential third cumulant is \(2\tau^3\) and a Gaussian has zero third cumulant, the standardized skewness above follows.^[1][3]

Two limits clarify membership. As \(\tau\to0\) (equivalently \(\lambda\to\infty\)), the exponential addend collapses to zero and \(Z\) approaches \(N(\mu,\sigma^2)\). As \(\sigma\to0\), \(X\) collapses to \(\mu\) and the limit is the shifted exponential \(\mu+Y\), whose support is \([\mu,\infty)\). These are limiting laws; at every positive \(\sigma\), the support remains all real numbers.^[1]

Knowledge Transfer

The construction travels unchanged between chromatographic peak fitting and reaction-time distribution fitting: a symmetric Gaussian term is convolved with a one-sided exponential term, and the same \(\mu,\sigma,\tau\) algebra describes location, spread and skew. What does not transfer is a causal interpretation. A chromatographic zone's tail and a participant's slow responses are not the same process simply because the same curve fits both.^[6][2][4]

The general mathematical idea of combining symmetric variation with one-sided delay could motivate other models, but the exact named EMG family requires the specified component laws and independence. The portable generic skeleton is a possible future-prime question; this entry remains a specific probability-distribution identity.^[1]

Examples

Capillary chromatography. The 2003 original study separates C10–C22 saturated fatty acids by reversed-phase capillary liquid chromatography and fits the resulting profiles with EMG functions. It compares fitted symmetric and asymmetric contributions with measured statistical moments, while changes in integration limits and signal-to-noise affect that comparison.^[6] Mapped back: Gaussian component = symmetric model part of a zone profile; exponential component = modeled one-sided tail scale \(\tau\); independence and sum = the EMG fitting assumption; support/right tail = mathematical full-real support and positive observed peak asymmetry, not literal negative elution time; parameter readout = \(\sigma\) and \(\tau\) compared with zone moments.

Stroop response times. Heathcote and colleagues fit normal-exponential convolutions to response-time distributions under Stroop conditions. Their distributional analysis found facilitation not evident from mean RT alone. The fitted tail parameter distinguishes shapes, but by itself does not diagnose a particular mental process.^[2][4] Mapped back: Gaussian component = fitted \(\mu,\sigma\) part of RT shape; exponential component = long-RT scale \(\tau\); independence and sum = mathematical fit construction; support/right tail = right-skewed times despite a theoretical negative support tail; parameter readout = condition comparisons without unique process labels.

Structural Tensions

Shape decomposition versus causal overinterpretation. The two-component form makes empirical skew easier to describe and can inspire a process hypothesis. But equating the fitted mathematical components with unique physical or cognitive stages risks false causal claims; treating the fit as wholly descriptive avoids that mistake while foregoing a mechanistic explanation.^[6][4] Diagnostic: Is there independent evidence that the two actual generators correspond to the fitted normal and exponential terms, or is only a shape comparison justified?

Tail preservation versus contamination control. Retaining long observations preserves valid evidence about \(\tau\); deleting them as outliers can bias the fitted right tail. Yet genuine artifacts or miscoded observations can destabilize a fit if retained. One cannot optimize both by a blanket rule to keep or trim extremes.^[3] Diagnostic: Which observations are independently invalid, and how do fitted parameters change under explicit truncation or contamination checks?

Structural–Framed Character

EMG sits toward the structural end of the structural–framed spectrum. Its normal-plus-exponential sum, convolution, moment identities and limiting laws are exact mathematical structure. The domain frame—probability distributions fitted to measured variables—remains constitutive, so a generic claim about “one-sided delay” is not by itself this identity.^[1]

The five transfer tests are met with boundaries: (1) Mechanism is independent addition; (2) Invariants are its density and moment relations; (3) Parameter freedom lies in \(\mu,\sigma,\lambda\); (4) Cross-setting transfer occurs between chromatographic peaks and RT distributions; (5) Boundary preservation keeps the mathematical law separate from application-specific causes and physically constrained support. Its character: a highly portable but typed statistical distribution, not a prime abstraction of delay itself.^[1][6][2]

Structural Core vs. Domain Accent

The core is the law of an independent Gaussian variable plus an exponential variable, with rate/scale conventions, full-real support when \(\sigma>0\), and right-skewed moments. The accents are chromatographic peaks, Stroop trials, chemical substances, and hypotheses about retention or cognition. Removing those accents leaves the distribution intact; removing Gaussian/exponential independence does not.^[1][6][2]

The live Probability Distribution node is the strict genus; Convolution supplies the necessary operation for its density. The broader “symmetric variability plus one-sided delay” skeleton is a future-prime question, not a second admitted identity here. No edge is inferred to the lexically similar Normal-Exponential-Gamma Distribution.^[1]

Instantiates / Related Primes

This entry presupposes Convolution and is a kind of Probability Distribution.

The proposed typed DAG uses strict subsumption under Probability Distribution and strict presupposition of Convolution. The latter is not a claim that every convolution is EMG; this named density is one particular convolution of two specified independent component densities. These relations await independent review and are not canonical edits.^[1]

The normal and exponential limits are relationships between families, not synonyms for the interior \(\sigma,\lambda>0\) law. Parameter limits change the distribution, and a skewed histogram can resemble EMG without satisfying its generative definition.^[1]

Relationships to Other Abstractions

Current abstraction Exponentially Modified Gaussian Distribution Domain-specific

Parents (2) — more general patterns this builds on

• Exponentially Modified Gaussian Distribution is a kind of Probability Distribution Domain-specific

  This independent normal-plus-exponential law is one parametric probability-distribution family.

• Exponentially Modified Gaussian Distribution presupposes Convolution Prime

  The density of the sum is defined as the convolution of its independent Gaussian and exponential component densities.

Hierarchy paths (6) — routes to 3 parentless roots

• Exponentially Modified Gaussian Distribution → Probability Distribution → Random Variable → Function (Mapping)

Show alternative paths (5)

Neighborhood in Abstraction Space

Exponentially Modified Gaussian Distribution sits in a sparse region of the domain-specific corpus (67^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Foundations of Probability & Inference (29 abstractions)

Nearest neighbors

• Cramér's Theorem (Large Deviations) — 0.85
• Covariance Matrix — 0.85
• Probability Distribution — 0.84
• Cochran's Theorem — 0.84
• Random Variable — 0.84

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With

• An arbitrary right-skewed curve: shape alone does not establish the exact convolution law.^[1]
• Gaussian minus exponential: a negative exponential addend reverses the tail orientation; its Wikipedia redirect is kept as unresolved provenance, not an alias.
• Normal-Exponential-Gamma Distribution: a different family with additional mixing structure, not this two-term independent sum.
• A unique physical or cognitive stage decomposition: fitted \(\mu,\sigma,\tau\) do not identify causes without external evidence.^[4]
• A physically bounded timing law: mathematical EMG has full-real support for positive \(\sigma\) even when observed times are nonnegative.^[1]

References

[1] SciPy developers, scipy.stats.exponnorm, Notes on density, independent-sum construction, real support and \(K=1/(\sigma\lambda)\). registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t

[2] A. Heathcote, S. J. Popiel and D. J. K. Mewhort, “Analysis of Response Time Distributions: An Example Using the Stroop Task”, Psychological Bulletin 109 (1991), 340–347, original abstract. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g

[3] R. Ulrich and J. Miller, “Effects of Truncation on Reaction Time Analysis”, Journal of Experimental Psychology: General 123 (1994), Appendix eqs. (35)–(37) and truncation study; PDF access intermittent. registry ↩a ↩b ↩c ↩d ↩e

[4] D. Matzke and E.-J. Wagenmakers, “Psychological Interpretation of the Ex-Gaussian and Shifted Wald Parameters”, Psychonomic Bulletin & Review 16 (2009), original abstract. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i

[5] E. Grushka, “Characterization of Exponentially Modified Gaussian Peaks in Chromatography”, Analytical Chemistry 44 (1972), 1733–1738; publisher record/first-page access only. registry ↩

[6] “Additivity of Statistical Moments in the Exponentially Modified Gaussian Model of Chromatography”, Analytica Chimica Acta 478 (2003), 99–110, publisher abstract and indexed excerpts. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h